This takes about ten minutes to read, and the sources are listed at the bottom. When a number comes from a news roundup and the company hasn't published it on its own page, I say so.
The main story
Put 30 November on your calendar.
A post on OpenAI's developer forum, dated 30 October 2026, walks through the company's deprecations page. According to the writer, Evals became read-only on 31 October, and Agent Builder, the Evals dashboard and API, and the v1/prompts API all shut down on 30 November. ChatKit is listed as staying. That doesn't help you if your ChatKit screen runs a workflow built inside Agent Builder, because the workflow is what's going away.
This is a summary of that forum post. I didn't check OpenAI's deprecations page again while writing this, so if you build on these products, open the page and go by what it says.
If you don't build software yourself, you can still ask the person who looks after your tools: "Do we use Agent Builder, or anything with v1/prompts in the address?" Write down the answer. It's a lot easier to sort out now than in December.
Where this fits in the course
Module 4 is about choosing your tools. You leave it with a main app, a backup, and a clear idea of which tools are allowed to see your work. See the 18 modules.
Other news this week
26 October · OpenAI · price from a tracker
GPT-6.1 Terra is available now
RetroAI Tracker describes Terra as a fast model that can look at a screen and writes in a normal-sounding voice. It lists the API price at $1 per million input tokens and $5 per million output tokens, and says Terra is in ChatGPT Work and Codex on paid plans. Those prices are the tracker's, and I haven't confirmed them on OpenAI's pricing page.
What to do with it: if your office already pays for ChatGPT, try Terra on a reply you were going to write anyway, then check it before you send it. Modules 3 and 15 cover why that check matters.
28 October · Google · limited access
Gemini 4 Carbon is only open to a few groups so far
The Model Press, in its roundup for 24 to 30 October, says Google announced Gemini 4 Carbon on 28 October and called it its strongest model for cybersecurity and long tasks. Access starts with cyber defenders and then trusted partners, so most people can't use it yet.
What to do with it: nothing for now. Wait until it's open to you before you plan any work around it.
27 October · Google DeepMind
Genie 4 builds interactive scenes from text
RetroAI Tracker says DeepMind announced Genie 4 on 27 October. It makes interactive scenes with sound that you steer by typing. It's starting as a research preview for trusted testers.
What to do with it: watch a clip if you're curious. It won't change your workday for a while.
26 October · Anthropic
Anthropic filed to go public
The same tracker says Anthropic submitted a Form S-1 on 26 October and plans to list under the ticker ANT. I haven't read the filing. It changes who owns the company behind Claude, and the way you use Claude stays the same.
21 October · Anthropic · a quiet change
Some Claude API settings were removed
RetroAI Tracker says Claude API version 2026-10-21 removes the temperature, top_p and top_k settings from the Messages API. The older API version is still listed as supported.
What to do with it: if a saved setup depends on one of those settings, it may now be ignored. It's worth keeping your prompts in one place, with one person who knows where they are, so changes like this get noticed.
A court fined lawyers for hidden AI instructions
On 20 October 2026, Brazil's Superior Court of Justice said its First Panel fined lawyers 20,000 reais for hiding instructions to an AI inside a court filing. The court's news post says the text sat in a footer bar, in the same colour as the bar so nobody would see it. The panel called it an act against the dignity of justice and sent the case to the bar association in Paraná and to federal prosecutors.
An earlier case in a labour court, written up by Tom's Hardware from a post by Professor Vladimir Aras dated 13 May 2026, used white text on a white page. The court fined the two lawyers who signed the filing about $16,000, which Tom's Hardware says was 10 percent of the award. The other side hadn't even filed a response yet.
Two takeaways for regular work. Never hide instructions in documents you send to anyone. And when you paste a PDF into an AI chat, remember it can read text you can't see. Ask it to quote anything in the document that looks like an instruction, then look at the document yourself.
Where this fits in the course
Module 15 covers checking the output, and Module 17 covers when to leave AI out of a job. The free 28 minutes includes a short lesson on checking.
Things you can ignore this week
Arguments about where Gemini 4 Carbon ranks on leaderboards. Most people can't use it yet, and the public rankings disagree with each other.
Headlines about math results from an OpenAI model that isn't out yet. The tracker says outside groups still have to review the claims.
Any new tool that asks for your customer or member list "just to try it." Names, addresses and member numbers stay out. That's Module 16.
Try this prompt
This uses the four-part prompt from the free lessons: the job, who it's for, the shape and the background. Take any names out before you paste it in.
Job: Tell me what would break if Agent Builder, an evals dashboard, or a saved-prompt API shut down on 30 November.
Who it is for: me, and I am not an engineer. Plain language.
Shape: a checklist of five questions I can ask the person who runs our tools, then a one-paragraph note I could forward.
Background: we use [name your apps]. Do not invent our setup. If I did not give you a fact, write "ask" instead of guessing.
Then open today's site and run the same prompt there. Keep whichever checklist is clearer.
One last thing
A model you can open today is more useful to you than one you can only read about. The people who get good at this run the same four-part prompt on their own work every week and check the draft before it goes anywhere. If you'd like to practice that with help, that's what the course is for.